基于深度强化学习的异构云任务调度研究

Research on Heterogeneous Cloud Task Scheduling Based on Deep Reinforcement Learning

在云环境下,如何在兼顾应用性能和虚拟机资源利用率的前提下,将大规模、动态的工作负载分配给云服务器是一个重要的挑战.为了解决这个问题,提出一种智能云任务调度器,针对任务的动态未知性,使用K-means实现任务聚类,获取任务类型.在此基础上,为解决云环境中动态异构性问题提出了基于异构感知的深度确定性策略梯度优化调度方法CDDPG.最后在阿里巴巴集群数据跟踪下进行了实验.结果表明,该调度器在响应时间和VM实例间负载均衡方面都有所提升.

In the cloud environment,it is an important challenge to allocate large-scale and dynamic workloads to cloud servers while taking into account application performance and virtual machine resource utilization.In order to solve this problem,an intelligent cloud task scheduler is proposed.Aiming at the dynamic unknown of tasks,K-means is used to achieve task clustering and obtain task types.On this basis,in order to solve the problem of dynamic heterogeneity in the cloud environment,a deep deterministic policy gradient optimization scheduling method based on heterogeneous perception is proposed.Finally,an experiment was conducted under Alibaba cluster data tracking.The results show that the scheduler has improved response time and load balancing among VM instances.